The AI slowdown pledge is now a federal antitrust lawsuit

One week ago, Dario Amodei asked the AI industry to slow down, and Elon Musk, Sam Altman and Demis Hassabis agreed on X within hours. On Friday, four people who pay for ChatGPT, Claude, Grok and Gemini sued all four companies over it.

The complaint is a scrapbook of X posts

The case is Buist v. Anthropic PBC, filed in the Northern District of California, according to Bloomberg Law. The plaintiffs are two Florida attorneys, a California attorney and activist named Cheyenne Hunt, and a California resident. Nick Rowley of Trial Lawyers for Justice is lead counsel, and the suit asks to represent every paid subscriber in the country.

The legal theory is Section 1 of the Sherman Act, the law against competitors agreeing to restrict output. The complaint calls the pledge "a classic output-restricting cartel" and argues that subscribers now pay the same price for products that improve more slowly than competition would produce. It asks for treble damages, an injunction against any agreement on training compute, release timing or development pace, and a jury trial.

What makes it unusual is the evidence. There are no leaked emails. The exhibits are the essays and posts everyone read last week: Jakub Pachocki's September 6 essay naming coordination as an option, Amodei's September 12 essay, Musk's "Dario is right," Altman's agreement, and Hassabis calling it "the right path forward." The labs made their case in public, and the plaintiffs simply screenshotted it.

The labs saw this coming

WIRED reported on September 10 that OpenAI had already asked members of Congress whether a coordinated slowdown would be legal. Amodei's own essay suggested the government "issue a narrow waiver for certain kinds of safety conversations." Then on Tuesday, OpenAI policy chief Chris Lehane told reporters the three labs had been talking for weeks and did not need a waiver. Altman said the company would not wait for an antitrust exemption. Three days later, the complaint landed.

The lawyer's framing borrows the industry's own fear and turns it around. Rowley told the AP that AI "could kill us all" if safety is left to "private self-serving agreements" between for-profit companies. That echoes a September 9 post from former Anthropic engineer @hilbertspaess, who wrote that the people building AI "earnestly believe that it could kill us all by the end of the decade." The plaintiffs take the same premise and reach the opposite conclusion: if the risk is real, four CEOs do not get to write the rules. Hunt said as much on X, according to ANI, posting that subscribers "deserve real AI safety standards" rather than a private deal.

The one lab that refused is the one not being sued

On Tuesday, before anyone had filed anything, @finkd posted Meta's answer to Amodei. Mark Zuckerberg wrote that "every lab has the responsibility and incentive to move at the pace required" to train safely, and that Meta had delayed its Muse agent for months without asking rivals to match. His argument was market pressure. People will not use agents that ignore what they ask, so alignment is a competitive edge, and "any lab that doesn't focus on alignment will fall behind." Nvidia's Jensen Huang said much the same at Dreamforce this week, calling speed versus safety a false choice.

Meta is not a defendant. I doubt the plaintiffs' lawyers read Zuckerberg's post as a legal filing, but it now works like one. The company that refused to agree cannot be accused of agreeing.

I think the plaintiffs have a real hurdle. Four executives publicly endorsing the same essay is not obviously an agreement, and each lab can say it acted alone. The threat is discovery. If the case survives a motion to dismiss, the private conversations Lehane described become evidence. A bipartisan bill introduced this month to carve safety collaboration out of antitrust law is not law yet. None of the four companies had commented as of Friday. I am not a lawyer and this is not legal advice, but a courtroom fight over how fast AI is allowed to improve is a new thing.

What this means if you run a business

How fast models improve is now a question for a court as well as a lab. Whether the slowdown comes by pledge or by injunction, the models under your tools may improve on a schedule that a judge or a standards body helps set.

That argues for building on what already works. Measure an AI process by the hours it saves your team this month, not by what the next model might do. A phone-intake bot that answers after 5 p.m. and an invoice follow-up that goes out on day 31 do not depend on anyone winning a race, or losing one.

If you want to know which of your repeated tasks are worth automating with today's models, New Face Design offers a free 20 minute process audit. You keep the findings either way.